Ever noticed that the longer you work on something with ChatGPT, the worse it gets? The first answer is quite good. By the tenth, it’s gone completely off the rails.
You’re developing a business strategy, researching a market or writing an important report. The first response is promising. You refine the brief, add context, challenge a few assumptions and ask it to go deeper.
But somewhere along the way, it starts contradicting earlier findings, forgetting key instructions and reintroducing things you’ve already ruled out. Each revision fixes one problem and creates another.
An hour later, you’re spending more time checking its work than actually making progress.
New research from Microsoft and Salesforce shows AI performance drops 39% when you brief it piece by piece. Here’s what’s going on and how to fix it.
I’m currently researching large language models for a data intelligence client, where we are benchmarking their specialist entity-matching platform against general-purpose AI, so when a paper explains why some AI conversations drift off course, I pay close attention. This one, LLMs Get Lost in Multi-Turn Conversation, received an Outstanding Paper Award at ICLR 2026, one of the most respected conferences in AI research, where it was one of only two papers honoured from around 19,000 submissions.
The researchers ran more than 200,000 simulated conversations across 15 leading models, including those behind ChatGPT and Claude. Each task was tested twice: once with everything the model needed in a single message, and once with the same information released gradually over several messages, which is how most of us actually work with AI. When the information arrived in instalments, performance fell by 39% on average.
How you brief AI will determine how good the output is.
The drop appeared even in conversations of just two messages, which points squarely at the quality of the brief. When AI doesn’t have the full picture, it fills the gaps with its own assumptions and hurries towards an answer, and when you later add the detail you’d forgotten, it holds on to its first guess and builds on top of it.
To sum it up: when AI takes a wrong turn in a conversation, it gets lost and does not recover.
Although the models’ underlying ability fell by only around 16%, their unreliability more than doubled. Running the same drip-fed conversation twice could produce very different results, with an average gap of 50 points out of 100 between the best and worst outcomes.
The strongest models lost roughly as much ground as the smaller ones, and models designed to reason before answering fared no better.
Answers grew longer as conversations went on, because each model kept carrying its earlier assumptions forward, and final answers ran anywhere from 20% to 300% longer than in the single-message versions.
Rushing proved costly as well. On the coding and maths tasks, when a model attempted a full answer early in the conversation, it scored less than half of what it achieved when it waited until it had all the information.
The models also gave the most weight to what was said first and last, so details shared in the middle of a conversation tended to fade from view.
The most telling result came when the researchers handed over exactly the same pieces of information in a single message, as a simple bullet list. The models retained about 95% of their performance, which shows that the content of the brief was sound and the damage came entirely from the way it was delivered.
The brief is the work!
This made me smile, because it’s exactly what I tell every client about briefing: brief in dribs and drabs and you’ll get dribs and drabs back. If anything, AI is less forgiving. A capable person will stop and ask a question when something is unclear, whereas AI usually makes a guess and then commits to it.
Here’s the working fix to this problem, tried and tested: When a project starts getting complicated, I stop asking ChatGPT to keep revising the same conversation. Instead, I ask it to summarise everything we’ve established: the objective, key findings, decisions made, constraints and anything we’ve ruled out. I review that summary, correct any mistakes, and start a fresh chat with it as the brief.
It gives the AI a clean starting point without losing the thinking we’ve already done. It’s a simple change that’s made a very noticeable difference to the quality of my work.
1. Write the brief before you open the chat. I take my time, and write a document first, covering what I want, who it’s for, the background, what ‘good’ looks like and anything it must avoid, all in one message. If you’re unsure where to begin, this template covers the essentials and takes about five minutes to fill in:
Task: what you want produced
Audience: who will read or use it
Context: the background, what has already been tried and any constraints
What good looks like: length, tone, format, and an example if you have one
Avoid: anything that is off limits
2. Ask it to ask. The research shows that AI rarely stops to check what it doesn’t know, so I build that step into the brief.
Ending with “Before you start, ask me any questions you need answered to do this well” turns its habit of guessing into a short round of questions, which you can answer in one reply before any work begins.
3. Hold back the answer until it has everything. If you need to share material over several messages, say so from the beginning:
“I’m going to send you several pieces of information. Don’t draft anything until I tell you I’m done.”
Early answer attempts were among the most costly mistakes in the study, and this single instruction prevents them.
4. Put the most important information, or what matters the most, at the start or the end. Because detail in the middle of a conversation tends to fade, I lead with the essentials and repeat any non-negotiables in the closing lines of a long brief.
5. Learn the signs that a chat has lost its way.
Stop correcting and start a new chat with the full brief if you notice any of these:
- Every answer is longer than the last. You asked for a tweak and got the whole thing back, bigger. (Research: the paper calls this “answer bloat”. Final answers ran 20% to 300% longer than one-message versions.)
- A mistake you fixed keeps coming back. You corrected it two rounds ago and it’s crept back in, because the AI is building on its earlier wrong answer, not your correction. (Research)
- It’s working from an assumption you never made. It guessed early to fill a gap in your brief and now treats that guess as fact. (Research)
- It forgets what you said in the middle. It remembers your first message and your latest one, but the detail you gave halfway through has vanished. (Research)
- It reacts only to your last message. You ask for one change and it quietly drops an earlier requirement to make it fit. (Research)
- It jumped to a full answer before you’d finished briefing. If it produced a complete draft after your first message, everything since has probably been patching that draft. (Research: early answers scored less than half as well.)
- Your new point gets bolted on, not worked in. The structure stays the same and your addition sits awkwardly on top.
- It apologises, then makes the same mistake again. “You’re absolutely right” followed by a near-identical answer is a clear sign it’s stuck.
- It suggests something you already said no to. Rejected ideas, headlines or directions start resurfacing.
- The tone or voice is drifting. It sounded like you at the start, and now it sounds like generic marketing copy.
- It’s stopped asking questions. A good collaborator asks when something’s unclear. If it just keeps guessing, it’s filling gaps with assumptions.
- You’re spending more time correcting than it would take to rebrief. That’s your cue. A fresh chat with a complete brief is almost always faster.
When you spot two or more, don’t argue with it. Ask it to summarise everything you’ve agreed into one brief, check that summary yourself, then paste it into a new chat. That’s what the researchers recommend: consolidate, then start again.
The “check that summary yourself” step matters, because a lost chat can carry its wrong assumptions into the summary too.
6. Consolidate before you move on. I ask the chat to pull everything together, then take that summary into a fresh conversation. This is the prompt I use:
Summarise everything important from this conversation into one brief I can paste into a new chat. Include every decision we’ve made and every detail you’d need.
Read the summary before you paste it, because a lost conversation can carry its wrong assumptions straight into the new one.
7. Build it once for anything you repeat.
If you do the same kind of task every week, like newsletters, content plans, proposals or client reports, stop briefing from scratch. Write one good brief and save it, so every new chat starts with the full picture and nothing left out. It’s the drip-feeding fix done once.
What goes into a reusable brief:
- Who you are and who it’s for. Your business, your audience and what they care about.
- The job. What this output is, what it’s for and what a great one achieves.
- Your voice. How you sound, words you use, words you never use, and spelling (UK or US).
- The format. Length, structure, headings, and where the call to action goes.
- Your rules. Things it must always do or never do: no invented stats, no em dashes, cite sources, whatever matters to you.
- An example of good. Paste in your best past version. One strong example does more than a page of instructions.
- What to ask first. Tell it to ask you for this week’s specifics before it starts writing, so it doesn’t guess.
The fastest way to build one: when a chat has gone well, ask it to write the brief for you before you close it:
Turn everything we’ve worked out in this chat into a reusable brief for this task. Include my audience, voice, format, rules and what a great result looks like, written so a fresh chat can follow it with no other context.
Read what it gives you and fix anything it got wrong. Then save it.
Where to save it:
- Claude: a Project (the brief goes in the project instructions, examples go in as files) or a Skill, which you can call up in any chat.
- ChatGPT: a Project, or a custom GPT if you want something you can share with your team.
- Gemini: a Gem.
- Any tool: a simple document you paste in at the start of a chat works too. The brief is what matters, not where it lives.
Keep it working:
- Test it in a fresh chat before you rely on it. If the first result needs heavy correcting, the brief is missing something.
- Update it when you correct something twice. A repeated correction is a gap in the brief.
- Keep the standing stuff in the brief and the weekly specifics in the chat. The brief covers what never changes. Each chat brings this week’s topic, data or news.
- Keep it tight. A one-page brief that’s clear beats five pages it has to dig through.
Some honesty…
The study used simulated users, and the researchers believe this makes their figures an underestimate, since real people wander off topic and give up when they become frustrated. The models tested were the leaders of May 2025 and the field has moved on considerably since then. The lesson still stands, because it concerns the way we brief, and that applies whichever model you happen to be using.
AI is a brilliant collaborator when you treat it as one, which means giving it the whole picture from the outset and, when it does lose its way, resisting the urge to follow it further into the woods and opening a fresh chat instead.
Credit Source: Mavgpt.ai
Monday Friday elsewhere